Rupanshu Soi

Standard-issue PhD student @ Stanford CS

[Rupanshu Soi @rupanshusoi]

Ears down bf, ears up gf

[Eggroll 🐻‍❄️ @EggrolarBear]

if you were a bunny would you be ears up bunny or ears down bunny

[Rupanshu Soi @rupanshusoi]

@xuanalogue Sounds like the categorical imperative

[Rupanshu Soi @rupanshusoi]

I feel Tanguy is even better, this one was painted in 1945

[Visakan Veerasamy @visakanv]

@eshear Alex Colville’s paintings in the 50s-70s look like they were inspired by 90s computer graphics

[Rupanshu Soi @rupanshusoi]

I recommend a few pages of Proust's magnum opus 'In Search of Lost Time' daily to wash away any and all influence of LLMs' mediocre prose-styling from your palate

[François Chollet @fchollet]

While the writing style of LLMs is still as recognizable as ever, a new trend is that humans have started organically writing like them, too (which makes sense: of course you would end up imitating the style you are constantly reading). That makes telling the difference between humans and clankers a bit more challenging.

[Rupanshu Soi @rupanshusoi]

How to see art (or pretty much anything else): https://t.co/qizg7Ktlab

[Rupanshu Soi @rupanshusoi]

If you would please consult the charts

[Rupanshu Soi @rupanshusoi]

I must regretfully inform you that property testing and property-based testing have nothing whatsoever to do with each other

[Rupanshu Soi @rupanshusoi]

Ramblings about programming systems of the future: https://t.co/WJi7s0cX7R

[Rupanshu Soi @rupanshusoi]

I always liked this anecdote:

> After Hilbert was told that a student in his class had dropped mathematics in order to become a poet, he is reported to have said "Good - he did not have enough imagination to become a mathematician"

Turns out poetry is harder 😭

[Rupanshu Soi @rupanshusoi]

Me and reviewer 2

[Rupanshu Soi @rupanshusoi]

Bro they even got the name of their own technique wrong?! The post calls it subquadratic sparse attn in one place and subquadratic selective attn in another 😭

[typedfemale @typedfemale]

their blogpost is 93% slop

[Rupanshu Soi @rupanshusoi]

@dog_rates By density you mean

[Rupanshu Soi @rupanshusoi]

That you have to swizzle the input matrices to a tensor core in some godforsaken way was just described to me as the leakiest abstraction in the world, as it is the floorplan of the chip leaking into the high-level programming model. Ha!

[Rupanshu Soi @rupanshusoi]

@itsclivetime So...VLIWs? Bring back trace scheduling!

[Rupanshu Soi @rupanshusoi]

Mike has been cooking this for a while, great to see it finally released!

[Michael Y. Li @michaelyli_]

Can a language model learn, end-to-end, what to keep in its own KV cache and what to throw away? Can it learn to forget while it learns to reason?

Deep learning's central lesson: capability emerges from end-to-end optimization, not heuristics/strong inductive biases. But for efficiency, we rely heavily on hand-designed approaches.

🗑️ Introducing Neural Garbage Collection (NGC): we train a language model to jointly reason and manage its own KV cache, using reinforcement learning with outcome-based task reward alone. No SFT, no proxy objectives, no summarization in natural language.

New paper with @jubayer_hamid, Emily Fox, and @noahdgoodman!

[Rupanshu Soi @rupanshusoi]

@heyyalexwang @michaelylin @jubayer_hamid @efoxdcs @noahdgoodman Whoops you tagged the wrong Michael

[Rupanshu Soi @rupanshusoi]

@allgarbled I wrote about this sometime ago! But I assumed it was a PhD thing as opposed to an SF / Bay area thing, but it very well could be: https://t.co/fjwf14GtWn

[Rupanshu Soi @rupanshusoi]

Yup, kind of counter intuitive at first, but the context is richer than your head

[François Chollet @fchollet]

You cannot think your way to a perfect design. Only building and testing, over many iterations, can reveal the flaws in your mental model and provide the feedback you need to create the best design possible.

[Rupanshu Soi @rupanshusoi]

Unfortunately, this is completely true. The amount of slop being produced with alacrity boggles the mind. And we simply don't have the human expertise or incentive structures in place to combat it

[Fernando Borretti @zetalyrae]

Many people, perhaps most, are made worse by AI. They're not "augmented", they just 10x their output of shit.